xbrz.cpp

SUBMITTED BY: Guest

DATE: Jan. 20, 2014, 5:36 p.m.

FORMAT: C++

SIZE: 46.5 kB

Raw Download

HITS: 982

Report
  1. // ****************************************************************************
  2. // * This file is part of the HqMAME project. It is distributed under *
  3. // * GNU General Public License: http://www.gnu.org/licenses/gpl.html *
  4. // * Copyright (C) Zenju (zenju AT gmx DOT de) - All Rights Reserved *
  5. // * *
  6. // * Additionally and as a special exception, the author gives permission *
  7. // * to link the code of this program with the MAME library (or with modified *
  8. // * versions of MAME that use the same license as MAME), and distribute *
  9. // * linked combinations including the two. You must obey the GNU General *
  10. // * Public License in all respects for all of the code used other than MAME. *
  11. // * If you modify this file, you may extend this exception to your version *
  12. // * of the file, but you are not obligated to do so. If you do not wish to *
  13. // * do so, delete this exception statement from your version. *
  14. // ****************************************************************************
  16. #include "xbrz.h"
  17. #include <cassert>
  18. #include <algorithm>
  20. namespace
  21. {
  22. template <uint32_t N> inline
  23. unsigned char getByte(uint32_t val) { return static_cast<unsigned char>((val >> (8 * N)) & 0xff); }
  25. inline unsigned char getRed (uint32_t val) { return getByte<2>(val); }
  26. inline unsigned char getGreen(uint32_t val) { return getByte<1>(val); }
  27. inline unsigned char getBlue (uint32_t val) { return getByte<0>(val); }
  29. template <class T> inline
  30. T abs(T value)
  31. {
  32. static_assert(std::is_signed<T>::value, "");
  33. return value < 0 ? -value : value;
  34. }
  36. const uint32_t redMask = 0xff0000;
  37. const uint32_t greenMask = 0x00ff00;
  38. const uint32_t blueMask = 0x0000ff;
  40. template <unsigned int N, unsigned int M> inline
  41. void alphaBlend(uint32_t& dst, uint32_t col) //blend color over destination with opacity N / M
  42. {
  43. static_assert(N < 256, "possible overflow of (col & redMask) * N");
  44. static_assert(M < 256, "possible overflow of (col & redMask ) * N + (dst & redMask ) * (M - N)");
  45. static_assert(0 < N && N < M, "");
  46. dst = (redMask & ((col & redMask ) * N + (dst & redMask ) * (M - N)) / M) | //this works because 8 upper bits are free
  47. (greenMask & ((col & greenMask) * N + (dst & greenMask) * (M - N)) / M) |
  48. (blueMask & ((col & blueMask ) * N + (dst & blueMask ) * (M - N)) / M);
  49. }
  52. //inline
  53. //double fastSqrt(double n)
  54. //{
  55. // __asm //speeds up xBRZ by about 9% compared to std::sqrt which internally uses the same assembler instructions but adds some "fluff"
  56. // {
  57. // fld n
  58. // fsqrt
  59. // }
  60. //}
  61. //
  64. inline
  65. uint32_t alphaBlend2(uint32_t pix1, uint32_t pix2, double alpha)
  66. {
  67. return (redMask & static_cast<uint32_t>((pix1 & redMask ) * alpha + (pix2 & redMask ) * (1 - alpha))) |
  68. (greenMask & static_cast<uint32_t>((pix1 & greenMask) * alpha + (pix2 & greenMask) * (1 - alpha))) |
  69. (blueMask & static_cast<uint32_t>((pix1 & blueMask ) * alpha + (pix2 & blueMask ) * (1 - alpha)));
  70. }
  73. uint32_t* byteAdvance( uint32_t* ptr, int bytes) { return reinterpret_cast< uint32_t*>(reinterpret_cast< char*>(ptr) + bytes); }
  74. const uint32_t* byteAdvance(const uint32_t* ptr, int bytes) { return reinterpret_cast<const uint32_t*>(reinterpret_cast<const char*>(ptr) + bytes); }
  77. //fill block with the given color
  78. inline
  79. void fillBlock(uint32_t* trg, int pitch, uint32_t col, int blockWidth, int blockHeight)
  80. {
  81. //for (int y = 0; y < blockHeight; ++y, trg = byteAdvance(trg, pitch))
  82. // std::fill(trg, trg + blockWidth, col);
  84. for (int y = 0; y < blockHeight; ++y, trg = byteAdvance(trg, pitch))
  85. for (int x = 0; x < blockWidth; ++x)
  86. trg[x] = col;
  87. }
  89. inline
  90. void fillBlock(uint32_t* trg, int pitch, uint32_t col, int n) { fillBlock(trg, pitch, col, n, n); }
  93. #ifdef _MSC_VER
  94. #define FORCE_INLINE __forceinline
  95. #elif defined __GNUC__
  96. #define FORCE_INLINE __attribute__((always_inline)) inline
  97. #else
  98. #define FORCE_INLINE inline
  99. #endif
  102. enum RotationDegree //clock-wise
  103. {
  104. ROT_0,
  105. ROT_90,
  106. ROT_180,
  107. ROT_270
  108. };
  110. //calculate input matrix coordinates after rotation at compile time
  111. template <RotationDegree rotDeg, size_t I, size_t J, size_t N>
  112. struct MatrixRotation;
  114. template <size_t I, size_t J, size_t N>
  115. struct MatrixRotation<ROT_0, I, J, N>
  116. {
  117. static const size_t I_old = I;
  118. static const size_t J_old = J;
  119. };
  121. template <RotationDegree rotDeg, size_t I, size_t J, size_t N> //(i, j) = (row, col) indices, N = size of (square) matrix
  122. struct MatrixRotation
  123. {
  124. static const size_t I_old = N - 1 - MatrixRotation<static_cast<RotationDegree>(rotDeg - 1), I, J, N>::J_old; //old coordinates before rotation!
  125. static const size_t J_old = MatrixRotation<static_cast<RotationDegree>(rotDeg - 1), I, J, N>::I_old; //
  126. };
  129. template <size_t N, RotationDegree rotDeg>
  130. class OutputMatrix
  131. {
  132. public:
  133. OutputMatrix(uint32_t* out, int outWidth) : //access matrix area, top-left at position "out" for image with given width
  134. out_(out),
  135. outWidth_(outWidth) {}
  137. template <size_t I, size_t J>
  138. uint32_t& ref() const
  139. {
  140. static const size_t I_old = MatrixRotation<rotDeg, I, J, N>::I_old;
  141. static const size_t J_old = MatrixRotation<rotDeg, I, J, N>::J_old;
  142. return *(out_ + J_old + I_old * outWidth_);
  143. }
  145. private:
  146. uint32_t* out_;
  147. const int outWidth_;
  148. };
  151. template <class T> inline
  152. T square(T value) { return value * value; }
  155. /*
  156. inline
  157. void rgbtoLuv(uint32_t c, double& L, double& u, double& v)
  158. {
  159. //http://www.easyrgb.com/index.php?X=MATH&H=02#text2
  160. double r = getRed (c) / 255.0;
  161. double g = getGreen(c) / 255.0;
  162. double b = getBlue (c) / 255.0;
  164. if ( r > 0.04045 )
  165. r = std::pow(( ( r + 0.055 ) / 1.055 ) , 2.4);
  166. else
  167. r /= 12.92;
  168. if ( g > 0.04045 )
  169. g = std::pow(( ( g + 0.055 ) / 1.055 ) , 2.4);
  170. else
  171. g /= 12.92;
  172. if ( b > 0.04045 )
  173. b = std::pow(( ( b + 0.055 ) / 1.055 ) , 2.4);
  174. else
  175. b /= 12.92;
  177. r *= 100;
  178. g *= 100;
  179. b *= 100;
  181. double x = 0.4124564 * r + 0.3575761 * g + 0.1804375 * b;
  182. double y = 0.2126729 * r + 0.7151522 * g + 0.0721750 * b;
  183. double z = 0.0193339 * r + 0.1191920 * g + 0.9503041 * b;
  184. //---------------------
  185. double var_U = 4 * x / ( x + 15 * y + 3 * z );
  186. double var_V = 9 * y / ( x + 15 * y + 3 * z );
  187. double var_Y = y / 100;
  189. if ( var_Y > 0.008856 ) var_Y = std::pow(var_Y , 1.0/3 );
  190. else var_Y = 7.787 * var_Y + 16.0 / 116;
  192. const double ref_X = 95.047; //Observer= 2°, Illuminant= D65
  193. const double ref_Y = 100.000;
  194. const double ref_Z = 108.883;
  196. const double ref_U = ( 4 * ref_X ) / ( ref_X + ( 15 * ref_Y ) + ( 3 * ref_Z ) );
  197. const double ref_V = ( 9 * ref_Y ) / ( ref_X + ( 15 * ref_Y ) + ( 3 * ref_Z ) );
  199. L = ( 116 * var_Y ) - 16;
  200. u = 13 * L * ( var_U - ref_U );
  201. v = 13 * L * ( var_V - ref_V );
  202. }
  203. */
  205. inline
  206. void rgbtoLab(uint32_t c, unsigned char& L, signed char& A, signed char& B)
  207. {
  208. //code: http://www.easyrgb.com/index.php?X=MATH
  209. //test: http://www.workwithcolor.com/color-converter-01.htm
  210. //------RGB to XYZ------
  211. double r = getRed (c) / 255.0;
  212. double g = getGreen(c) / 255.0;
  213. double b = getBlue (c) / 255.0;
  215. r = r > 0.04045 ? std::pow(( r + 0.055 ) / 1.055, 2.4) : r / 12.92;
  216. r = g > 0.04045 ? std::pow(( g + 0.055 ) / 1.055, 2.4) : g / 12.92;
  217. r = b > 0.04045 ? std::pow(( b + 0.055 ) / 1.055, 2.4) : b / 12.92;
  219. r *= 100;
  220. g *= 100;
  221. b *= 100;
  223. double x = 0.4124564 * r + 0.3575761 * g + 0.1804375 * b;
  224. double y = 0.2126729 * r + 0.7151522 * g + 0.0721750 * b;
  225. double z = 0.0193339 * r + 0.1191920 * g + 0.9503041 * b;
  226. //------XYZ to Lab------
  227. const double refX = 95.047; //
  228. const double refY = 100.000; //Observer= 2°, Illuminant= D65
  229. const double refZ = 108.883; //
  230. double var_X = x / refX;
  231. double var_Y = y / refY;
  232. double var_Z = z / refZ;
  234. var_X = var_X > 0.008856 ? std::pow(var_X, 1.0 / 3) : 7.787 * var_X + 4.0 / 29;
  235. var_Y = var_Y > 0.008856 ? std::pow(var_Y, 1.0 / 3) : 7.787 * var_Y + 4.0 / 29;
  236. var_Z = var_Z > 0.008856 ? std::pow(var_Z, 1.0 / 3) : 7.787 * var_Z + 4.0 / 29;
  238. L = static_cast<unsigned char>(116 * var_Y - 16);
  239. A = static_cast< signed char>(500 * (var_X - var_Y));
  240. B = static_cast< signed char>(200 * (var_Y - var_Z));
  241. };
  244. inline
  245. double distLAB(uint32_t pix1, uint32_t pix2)
  246. {
  247. unsigned char L1 = 0; //[0, 100]
  248. signed char a1 = 0; //[-128, 127]
  249. signed char b1 = 0; //[-128, 127]
  250. rgbtoLab(pix1, L1, a1, b1);
  252. unsigned char L2 = 0;
  253. signed char a2 = 0;
  254. signed char b2 = 0;
  255. rgbtoLab(pix2, L2, a2, b2);
  257. //-----------------------------
  258. //http://www.easyrgb.com/index.php?X=DELT
  260. //Delta E/CIE76
  261. return std::sqrt(square(1.0 * L1 - L2) +
  262. square(1.0 * a1 - a2) +
  263. square(1.0 * b1 - b2));
  264. }
  267. /*
  268. inline
  269. void rgbtoHsl(uint32_t c, double& h, double& s, double& l)
  270. {
  271. //http://www.easyrgb.com/index.php?X=MATH&H=18#text18
  272. const int r = getRed (c);
  273. const int g = getGreen(c);
  274. const int b = getBlue (c);
  276. const int varMin = numeric::min(r, g, b);
  277. const int varMax = numeric::max(r, g, b);
  278. const int delMax = varMax - varMin;
  280. l = (varMax + varMin) / 2.0 / 255.0;
  282. if (delMax == 0) //gray, no chroma...
  283. {
  284. h = 0;
  285. s = 0;
  286. }
  287. else
  288. {
  289. s = l < 0.5 ?
  290. delMax / (1.0 * varMax + varMin) :
  291. delMax / (2.0 * 255 - varMax - varMin);
  293. double delR = ((varMax - r) / 6.0 + delMax / 2.0) / delMax;
  294. double delG = ((varMax - g) / 6.0 + delMax / 2.0) / delMax;
  295. double delB = ((varMax - b) / 6.0 + delMax / 2.0) / delMax;
  297. if (r == varMax)
  298. h = delB - delG;
  299. else if (g == varMax)
  300. h = 1 / 3.0 + delR - delB;
  301. else if (b == varMax)
  302. h = 2 / 3.0 + delG - delR;
  304. if (h < 0)
  305. h += 1;
  306. if (h > 1)
  307. h -= 1;
  308. }
  309. }
  311. inline
  312. double distHSL(uint32_t pix1, uint32_t pix2, double lightningWeight)
  313. {
  314. double h1 = 0;
  315. double s1 = 0;
  316. double l1 = 0;
  317. rgbtoHsl(pix1, h1, s1, l1);
  318. double h2 = 0;
  319. double s2 = 0;
  320. double l2 = 0;
  321. rgbtoHsl(pix2, h2, s2, l2);
  323. //HSL is in cylindric coordinatates where L represents height, S radius, H angle,
  324. //however we interpret the cylinder as a bi-conic solid with top/bottom radius 0, middle radius 1
  325. assert(0 <= h1 && h1 <= 1);
  326. assert(0 <= h2 && h2 <= 1);
  328. double r1 = l1 < 0.5 ?
  329. l1 * 2 :
  330. 2 - l1 * 2;
  332. double x1 = r1 * s1 * std::cos(h1 * 2 * numeric::pi);
  333. double y1 = r1 * s1 * std::sin(h1 * 2 * numeric::pi);
  334. double z1 = l1;
  336. double r2 = l2 < 0.5 ?
  337. l2 * 2 :
  338. 2 - l2 * 2;
  340. double x2 = r2 * s2 * std::cos(h2 * 2 * numeric::pi);
  341. double y2 = r2 * s2 * std::sin(h2 * 2 * numeric::pi);
  342. double z2 = l2;
  344. return 255 * std::sqrt(square(x1 - x2) + square(y1 - y2) + square(lightningWeight * (z1 - z2)));
  345. }
  346. */
  349. inline
  350. double distRGB(uint32_t pix1, uint32_t pix2)
  351. {
  352. const double r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2);
  353. const double g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
  354. const double b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
  356. //euklidean RGB distance
  357. return std::sqrt(square(r_diff) + square(g_diff) + square(b_diff));
  358. }
  361. inline
  362. double distNonLinearRGB(uint32_t pix1, uint32_t pix2)
  363. {
  364. //non-linear rgb: http://www.compuphase.com/cmetric.htm
  365. const double r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2);
  366. const double g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
  367. const double b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
  369. const double r_avg = (static_cast<double>(getRed(pix1)) + getRed(pix2)) / 2;
  370. return std::sqrt((2 + r_avg / 255) * square(r_diff) + 4 * square(g_diff) + (2 + (255 - r_avg) / 255) * square(b_diff));
  371. }
  374. inline
  375. double distYCbCr(uint32_t pix1, uint32_t pix2, double lumaWeight)
  376. {
  377. //http://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion
  378. //YCbCr conversion is a matrix multiplication => take advantage of linearity by subtracting first!
  379. const int r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2); //we may delay division by 255 to after matrix multiplication
  380. const int g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2); //
  381. const int b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2); //substraction for int is noticeable faster than for double!
  383. const double k_b = 0.0722; //ITU-R BT.709 conversion
  384. const double k_r = 0.2126; //
  385. const double k_g = 1 - k_b - k_r;
  387. const double scale_b = 0.5 / (1 - k_b);
  388. const double scale_r = 0.5 / (1 - k_r);
  390. const double y = k_r * r_diff + k_g * g_diff + k_b * b_diff; //[!], analog YCbCr!
  391. const double c_b = scale_b * (b_diff - y);
  392. const double c_r = scale_r * (r_diff - y);
  394. //we skip division by 255 to have similar range like other distance functions
  395. return std::sqrt(square(lumaWeight * y) + square(c_b) + square(c_r));
  396. }
  399. inline
  400. double distYUV(uint32_t pix1, uint32_t pix2, double luminanceWeight)
  401. {
  402. //perf: it's not worthwhile to buffer the YUV-conversion, the direct code is faster by ~ 6%
  403. //since RGB -> YUV conversion is essentially a matrix multiplication, we can calculate the RGB diff before the conversion (distributive property)
  404. const double r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2);
  405. const double g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
  406. const double b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
  408. //http://en.wikipedia.org/wiki/YUV#Conversion_to.2Ffrom_RGB
  409. const double w_b = 0.114;
  410. const double w_r = 0.299;
  411. const double w_g = 1 - w_r - w_b;
  413. const double u_max = 0.436;
  414. const double v_max = 0.615;
  416. const double scale_u = u_max / (1 - w_b);
  417. const double scale_v = v_max / (1 - w_r);
  419. double y = w_r * r_diff + w_g * g_diff + w_b * b_diff;//value range: 255 * [-1, 1]
  420. double u = scale_u * (b_diff - y); //value range: 255 * 2 * u_max * [-1, 1]
  421. double v = scale_v * (r_diff - y); //value range: 255 * 2 * v_max * [-1, 1]
  423. #ifndef NDEBUG
  424. const double eps = 0.5;
  425. #endif
  426. assert(std::abs(y) <= 255 + eps);
  427. assert(std::abs(u) <= 255 * 2 * u_max + eps);
  428. assert(std::abs(v) <= 255 * 2 * v_max + eps);
  430. return std::sqrt(square(luminanceWeight * y) + square(u) + square(v));
  431. }
  434. inline
  435. double colorDist(uint32_t pix1, uint32_t pix2, double luminanceWeight)
  436. {
  437. if (pix1 == pix2) //about 8% perf boost
  438. return 0;
  440. //return distHSL(pix1, pix2, luminanceWeight);
  441. //return distRGB(pix1, pix2);
  442. //return distLAB(pix1, pix2);
  443. //return distNonLinearRGB(pix1, pix2);
  444. //return distYUV(pix1, pix2, luminanceWeight);
  446. return distYCbCr(pix1, pix2, luminanceWeight);
  447. }
  449. inline
  450. bool equalColor(uint32_t col1, uint32_t col2, double luminanceWeight, double equalColorTolerance)
  451. {
  452. return colorDist(col1, col2, luminanceWeight) < equalColorTolerance;
  453. }
  456. enum BlendType
  457. {
  458. BLEND_NONE = 0,
  459. BLEND_NORMAL, //a normal indication to blend
  460. BLEND_DOMINANT, //a strong indication to blend
  461. //attention: BlendType must fit into the value range of 2 bit!!!
  462. };
  464. struct BlendResult
  465. {
  466. BlendType
  467. /**/blend_f, blend_g,
  468. /**/blend_j, blend_k;
  469. };
  472. struct Kernel_4x4 //kernel for preprocessing step
  473. {
  474. uint32_t
  475. /**/a, b, c, d,
  476. /**/e, f, g, h,
  477. /**/i, j, k, l,
  478. /**/m, n, o, p;
  479. };
  481. /*
  482. input kernel area naming convention:
  483. -----------------
  484. | A | B | C | D |
  485. ----|---|---|---|
  486. | E | F | G | H | //evalute the four corners between F, G, J, K
  487. ----|---|---|---| //input pixel is at position F
  488. | I | J | K | L |
  489. ----|---|---|---|
  490. | M | N | O | P |
  491. -----------------
  492. */
  493. FORCE_INLINE //detect blend direction
  494. BlendResult preProcessCorners(const Kernel_4x4& ker, const xbrz::ScalerCfg& cfg) //result: F, G, J, K corners of "GradientType"
  495. {
  496. BlendResult result = {};
  498. if ((ker.f == ker.g &&
  499. ker.j == ker.k) ||
  500. (ker.f == ker.j &&
  501. ker.g == ker.k))
  502. return result;
  504. auto dist = [&](uint32_t col1, uint32_t col2) { return colorDist(col1, col2, cfg.luminanceWeight_); };
  506. const int weight = 4;
  507. double jg = dist(ker.i, ker.f) + dist(ker.f, ker.c) + dist(ker.n, ker.k) + dist(ker.k, ker.h) + weight * dist(ker.j, ker.g);
  508. double fk = dist(ker.e, ker.j) + dist(ker.j, ker.o) + dist(ker.b, ker.g) + dist(ker.g, ker.l) + weight * dist(ker.f, ker.k);
  510. if (jg < fk) //test sample: 70% of values max(jg, fk) / min(jg, fk) are between 1.1 and 3.7 with median being 1.8
  511. {
  512. const bool dominantGradient = cfg.dominantDirectionThreshold * jg < fk;
  513. if (ker.f != ker.g && ker.f != ker.j)
  514. result.blend_f = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
  516. if (ker.k != ker.j && ker.k != ker.g)
  517. result.blend_k = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
  518. }
  519. else if (fk < jg)
  520. {
  521. const bool dominantGradient = cfg.dominantDirectionThreshold * fk < jg;
  522. if (ker.j != ker.f && ker.j != ker.k)
  523. result.blend_j = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
  525. if (ker.g != ker.f && ker.g != ker.k)
  526. result.blend_g = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
  527. }
  528. return result;
  529. }
  531. struct Kernel_3x3
  532. {
  533. uint32_t
  534. /**/a, b, c,
  535. /**/d, e, f,
  536. /**/g, h, i;
  537. };
  539. #define DEF_GETTER(x) template <RotationDegree rotDeg> uint32_t inline get_##x(const Kernel_3x3& ker) { return ker.x; }
  540. //we cannot and NEED NOT write "ker.##x" since ## concatenates preprocessor tokens but "." is not a token
  541. DEF_GETTER(a) DEF_GETTER(b) DEF_GETTER(c)
  542. DEF_GETTER(d) DEF_GETTER(e) DEF_GETTER(f)
  543. DEF_GETTER(g) DEF_GETTER(h) DEF_GETTER(i)
  544. #undef DEF_GETTER
  546. #define DEF_GETTER(x, y) template <> inline uint32_t get_##x<ROT_90>(const Kernel_3x3& ker) { return ker.y; }
  547. DEF_GETTER(a, g) DEF_GETTER(b, d) DEF_GETTER(c, a)
  548. DEF_GETTER(d, h) DEF_GETTER(e, e) DEF_GETTER(f, b)
  549. DEF_GETTER(g, i) DEF_GETTER(h, f) DEF_GETTER(i, c)
  550. #undef DEF_GETTER
  552. #define DEF_GETTER(x, y) template <> inline uint32_t get_##x<ROT_180>(const Kernel_3x3& ker) { return ker.y; }
  553. DEF_GETTER(a, i) DEF_GETTER(b, h) DEF_GETTER(c, g)
  554. DEF_GETTER(d, f) DEF_GETTER(e, e) DEF_GETTER(f, d)
  555. DEF_GETTER(g, c) DEF_GETTER(h, b) DEF_GETTER(i, a)
  556. #undef DEF_GETTER
  558. #define DEF_GETTER(x, y) template <> inline uint32_t get_##x<ROT_270>(const Kernel_3x3& ker) { return ker.y; }
  559. DEF_GETTER(a, c) DEF_GETTER(b, f) DEF_GETTER(c, i)
  560. DEF_GETTER(d, b) DEF_GETTER(e, e) DEF_GETTER(f, h)
  561. DEF_GETTER(g, a) DEF_GETTER(h, d) DEF_GETTER(i, g)
  562. #undef DEF_GETTER
  565. //compress four blend types into a single byte
  566. inline BlendType getTopL (unsigned char b) { return static_cast<BlendType>(0x3 & b); }
  567. inline BlendType getTopR (unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 2)); }
  568. inline BlendType getBottomR(unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 4)); }
  569. inline BlendType getBottomL(unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 6)); }
  571. inline void setTopL (unsigned char& b, BlendType bt) { b |= bt; } //buffer is assumed to be initialized before preprocessing!
  572. inline void setTopR (unsigned char& b, BlendType bt) { b |= (bt << 2); }
  573. inline void setBottomR(unsigned char& b, BlendType bt) { b |= (bt << 4); }
  574. inline void setBottomL(unsigned char& b, BlendType bt) { b |= (bt << 6); }
  576. inline bool blendingNeeded(unsigned char b) { return b != 0; }
  578. template <RotationDegree rotDeg> inline
  579. unsigned char rotateBlendInfo(unsigned char b) { return b; }
  580. template <> inline unsigned char rotateBlendInfo<ROT_90 >(unsigned char b) { return ((b << 2) | (b >> 6)) & 0xff; }
  581. template <> inline unsigned char rotateBlendInfo<ROT_180>(unsigned char b) { return ((b << 4) | (b >> 4)) & 0xff; }
  582. template <> inline unsigned char rotateBlendInfo<ROT_270>(unsigned char b) { return ((b << 6) | (b >> 2)) & 0xff; }
  585. #ifndef NDEBUG
  586. int debugPixelX = -1;
  587. int debugPixelY = 84;
  588. bool breakIntoDebugger = false;
  589. #endif
  592. /*
  593. input kernel area naming convention:
  594. -------------
  595. | A | B | C |
  596. ----|---|---|
  597. | D | E | F | //input pixel is at position E
  598. ----|---|---|
  599. | G | H | I |
  600. -------------
  601. */
  602. template <class Scaler, RotationDegree rotDeg>
  603. FORCE_INLINE //perf: quite worth it!
  604. void scalePixel(const Kernel_3x3& ker,
  605. uint32_t* target, int trgWidth,
  606. unsigned char blendInfo, //result of preprocessing all four corners of pixel "e"
  607. const xbrz::ScalerCfg& cfg)
  608. {
  609. #define a get_a<rotDeg>(ker)
  610. #define b get_b<rotDeg>(ker)
  611. #define c get_c<rotDeg>(ker)
  612. #define d get_d<rotDeg>(ker)
  613. #define e get_e<rotDeg>(ker)
  614. #define f get_f<rotDeg>(ker)
  615. #define g get_g<rotDeg>(ker)
  616. #define h get_h<rotDeg>(ker)
  617. #define i get_i<rotDeg>(ker)
  619. #ifndef NDEBUG
  620. if (breakIntoDebugger)
  621. __debugbreak(); //__asm int 3;
  622. #endif
  624. const unsigned char blend = rotateBlendInfo<rotDeg>(blendInfo);
  626. if (getBottomR(blend) >= BLEND_NORMAL)
  627. {
  628. auto eq = [&](uint32_t col1, uint32_t col2) { return equalColor(col1, col2, cfg.luminanceWeight_, cfg.equalColorTolerance_); };
  629. auto dist = [&](uint32_t col1, uint32_t col2) { return colorDist(col1, col2, cfg.luminanceWeight_); };
  631. const bool doLineBlend = [&]() -> bool
  632. {
  633. if (getBottomR(blend) >= BLEND_DOMINANT)
  634. return true;
  636. //make sure there is no second blending in an adjacent rotation for this pixel: handles insular pixels, mario eyes
  637. if (getTopR(blend) != BLEND_NONE && !eq(e, g)) //but support double-blending for 90° corners
  638. return false;
  639. if (getBottomL(blend) != BLEND_NONE && !eq(e, c))
  640. return false;
  642. //no full blending for L-shapes; blend corner only (handles "mario mushroom eyes")
  643. if (eq(g, h) && eq(h , i) && eq(i, f) && eq(f, c) && !eq(e, i))
  644. return false;
  646. return true;
  647. }();
  649. const uint32_t px = dist(e, f) <= dist(e, h) ? f : h; //choose most similar color
  651. OutputMatrix<Scaler::scale, rotDeg> out(target, trgWidth);
  653. if (doLineBlend)
  654. {
  655. const double fg = dist(f, g); //test sample: 70% of values max(fg, hc) / min(fg, hc) are between 1.1 and 3.7 with median being 1.9
  656. const double hc = dist(h, c); //
  658. const bool haveShallowLine = cfg.steepDirectionThreshold * fg <= hc && e != g && d != g;
  659. const bool haveSteepLine = cfg.steepDirectionThreshold * hc <= fg && e != c && b != c;
  661. if (haveShallowLine)
  662. {
  663. if (haveSteepLine)
  664. Scaler::blendLineSteepAndShallow(px, out);
  665. else
  666. Scaler::blendLineShallow(px, out);
  667. }
  668. else
  669. {
  670. if (haveSteepLine)
  671. Scaler::blendLineSteep(px, out);
  672. else
  673. Scaler::blendLineDiagonal(px,out);
  674. }
  675. }
  676. else
  677. Scaler::blendCorner(px, out);
  678. }
  680. #undef a
  681. #undef b
  682. #undef c
  683. #undef d
  684. #undef e
  685. #undef f
  686. #undef g
  687. #undef h
  688. #undef i
  689. }
  692. template <class Scaler> //scaler policy: see "Scaler2x" reference implementation
  693. void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, const xbrz::ScalerCfg& cfg, int yFirst, int yLast)
  694. {
  695. yFirst = std::max(yFirst, 0);
  696. yLast = std::min(yLast, srcHeight);
  697. if (yFirst >= yLast || srcWidth <= 0)
  698. return;
  700. const int trgWidth = srcWidth * Scaler::scale;
  702. //"use" space at the end of the image as temporary buffer for "on the fly preprocessing": we even could use larger area of
  703. //"sizeof(uint32_t) * srcWidth * (yLast - yFirst)" bytes without risk of accidental overwriting before accessing
  704. const int bufferSize = srcWidth;
  705. unsigned char* preProcBuffer = reinterpret_cast<unsigned char*>(trg + yLast * Scaler::scale * trgWidth) - bufferSize;
  706. std::fill(preProcBuffer, preProcBuffer + bufferSize, 0);
  707. static_assert(BLEND_NONE == 0, "");
  709. //initialize preprocessing buffer for first row: detect upper left and right corner blending
  710. //this cannot be optimized for adjacent processing stripes; we must not allow for a memory race condition!
  711. if (yFirst > 0)
  712. {
  713. const int y = yFirst - 1;
  715. const uint32_t* s_m1 = src + srcWidth * std::max(y - 1, 0);
  716. const uint32_t* s_0 = src + srcWidth * y; //center line
  717. const uint32_t* s_p1 = src + srcWidth * std::min(y + 1, srcHeight - 1);
  718. const uint32_t* s_p2 = src + srcWidth * std::min(y + 2, srcHeight - 1);
  720. for (int x = 0; x < srcWidth; ++x)
  721. {
  722. const int x_m1 = std::max(x - 1, 0);
  723. const int x_p1 = std::min(x + 1, srcWidth - 1);
  724. const int x_p2 = std::min(x + 2, srcWidth - 1);
  726. Kernel_4x4 ker = {}; //perf: initialization is negligable
  727. ker.a = s_m1[x_m1]; //read sequentially from memory as far as possible
  728. ker.b = s_m1[x];
  729. ker.c = s_m1[x_p1];
  730. ker.d = s_m1[x_p2];
  732. ker.e = s_0[x_m1];
  733. ker.f = s_0[x];
  734. ker.g = s_0[x_p1];
  735. ker.h = s_0[x_p2];
  737. ker.i = s_p1[x_m1];
  738. ker.j = s_p1[x];
  739. ker.k = s_p1[x_p1];
  740. ker.l = s_p1[x_p2];
  742. ker.m = s_p2[x_m1];
  743. ker.n = s_p2[x];
  744. ker.o = s_p2[x_p1];
  745. ker.p = s_p2[x_p2];
  747. const BlendResult res = preProcessCorners(ker, cfg);
  748. /*
  749. preprocessing blend result:
  750. ---------
  751. | F | G | //evalute corner between F, G, J, K
  752. ----|---| //input pixel is at position F
  753. | J | K |
  754. ---------
  755. */
  756. setTopR(preProcBuffer[x], res.blend_j);
  758. if (x + 1 < srcWidth)
  759. setTopL(preProcBuffer[x + 1], res.blend_k);
  760. }
  761. }
  762. //------------------------------------------------------------------------------------
  764. for (int y = yFirst; y < yLast; ++y)
  765. {
  766. uint32_t* out = trg + Scaler::scale * y * trgWidth; //consider MT "striped" access
  768. const uint32_t* s_m1 = src + srcWidth * std::max(y - 1, 0);
  769. const uint32_t* s_0 = src + srcWidth * y; //center line
  770. const uint32_t* s_p1 = src + srcWidth * std::min(y + 1, srcHeight - 1);
  771. const uint32_t* s_p2 = src + srcWidth * std::min(y + 2, srcHeight - 1);
  773. unsigned char blend_xy1 = 0; //corner blending for current (x, y + 1) position
  775. for (int x = 0; x < srcWidth; ++x, out += Scaler::scale)
  776. {
  777. #ifndef NDEBUG
  778. breakIntoDebugger = debugPixelX == x && debugPixelY == y;
  779. #endif
  780. //all those bounds checks have only insignificant impact on performance!
  781. const int x_m1 = std::max(x - 1, 0); //perf: prefer array indexing to additional pointers!
  782. const int x_p1 = std::min(x + 1, srcWidth - 1);
  783. const int x_p2 = std::min(x + 2, srcWidth - 1);
  785. //evaluate the four corners on bottom-right of current pixel
  786. unsigned char blend_xy = 0; //for current (x, y) position
  787. {
  788. Kernel_4x4 ker = {}; //perf: initialization is negligable
  789. ker.a = s_m1[x_m1]; //read sequentially from memory as far as possible
  790. ker.b = s_m1[x];
  791. ker.c = s_m1[x_p1];
  792. ker.d = s_m1[x_p2];
  794. ker.e = s_0[x_m1];
  795. ker.f = s_0[x];
  796. ker.g = s_0[x_p1];
  797. ker.h = s_0[x_p2];
  799. ker.i = s_p1[x_m1];
  800. ker.j = s_p1[x];
  801. ker.k = s_p1[x_p1];
  802. ker.l = s_p1[x_p2];
  804. ker.m = s_p2[x_m1];
  805. ker.n = s_p2[x];
  806. ker.o = s_p2[x_p1];
  807. ker.p = s_p2[x_p2];
  809. const BlendResult res = preProcessCorners(ker, cfg);
  810. /*
  811. preprocessing blend result:
  812. ---------
  813. | F | G | //evalute corner between F, G, J, K
  814. ----|---| //current input pixel is at position F
  815. | J | K |
  816. ---------
  817. */
  818. blend_xy = preProcBuffer[x];
  819. setBottomR(blend_xy, res.blend_f); //all four corners of (x, y) have been determined at this point due to processing sequence!
  821. setTopR(blend_xy1, res.blend_j); //set 2nd known corner for (x, y + 1)
  822. preProcBuffer[x] = blend_xy1; //store on current buffer position for use on next row
  824. blend_xy1 = 0;
  825. setTopL(blend_xy1, res.blend_k); //set 1st known corner for (x + 1, y + 1) and buffer for use on next column
  827. if (x + 1 < srcWidth) //set 3rd known corner for (x + 1, y)
  828. setBottomL(preProcBuffer[x + 1], res.blend_g);
  829. }
  831. //fill block of size scale * scale with the given color
  832. fillBlock(out, trgWidth * sizeof(uint32_t), s_0[x], Scaler::scale); //place *after* preprocessing step, to not overwrite the results while processing the the last pixel!
  834. //blend four corners of current pixel
  835. if (blendingNeeded(blend_xy)) //good 20% perf-improvement
  836. {
  837. Kernel_3x3 ker = {}; //perf: initialization is negligable
  839. ker.a = s_m1[x_m1]; //read sequentially from memory as far as possible
  840. ker.b = s_m1[x];
  841. ker.c = s_m1[x_p1];
  843. ker.d = s_0[x_m1];
  844. ker.e = s_0[x];
  845. ker.f = s_0[x_p1];
  847. ker.g = s_p1[x_m1];
  848. ker.h = s_p1[x];
  849. ker.i = s_p1[x_p1];
  851. scalePixel<Scaler, ROT_0 >(ker, out, trgWidth, blend_xy, cfg);
  852. scalePixel<Scaler, ROT_90 >(ker, out, trgWidth, blend_xy, cfg);
  853. scalePixel<Scaler, ROT_180>(ker, out, trgWidth, blend_xy, cfg);
  854. scalePixel<Scaler, ROT_270>(ker, out, trgWidth, blend_xy, cfg);
  855. }
  856. }
  857. }
  858. }
  861. struct Scaler2x
  862. {
  863. static const int scale = 2;
  865. template <class OutputMatrix>
  866. static void blendLineShallow(uint32_t col, OutputMatrix& out)
  867. {
  868. alphaBlend<1, 4>(out.template ref<scale - 1, 0>(), col);
  869. alphaBlend<3, 4>(out.template ref<scale - 1, 1>(), col);
  870. }
  872. template <class OutputMatrix>
  873. static void blendLineSteep(uint32_t col, OutputMatrix& out)
  874. {
  875. alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col);
  876. alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col);
  877. }
  879. template <class OutputMatrix>
  880. static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
  881. {
  882. alphaBlend<1, 4>(out.template ref<1, 0>(), col);
  883. alphaBlend<1, 4>(out.template ref<0, 1>(), col);
  884. alphaBlend<5, 6>(out.template ref<1, 1>(), col); //[!] fixes 7/8 used in xBR
  885. }
  887. template <class OutputMatrix>
  888. static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
  889. {
  890. alphaBlend<1, 2>(out.template ref<1, 1>(), col);
  891. }
  893. template <class OutputMatrix>
  894. static void blendCorner(uint32_t col, OutputMatrix& out)
  895. {
  896. //model a round corner
  897. alphaBlend<21, 100>(out.template ref<1, 1>(), col); //exact: 1 - pi/4 = 0.2146018366
  898. }
  899. };
  902. struct Scaler3x
  903. {
  904. static const int scale = 3;
  906. template <class OutputMatrix>
  907. static void blendLineShallow(uint32_t col, OutputMatrix& out)
  908. {
  909. alphaBlend<1, 4>(out.template ref<scale - 1, 0>(), col);
  910. alphaBlend<1, 4>(out.template ref<scale - 2, 2>(), col);
  912. alphaBlend<3, 4>(out.template ref<scale - 1, 1>(), col);
  913. out.template ref<scale - 1, 2>() = col;
  914. }
  916. template <class OutputMatrix>
  917. static void blendLineSteep(uint32_t col, OutputMatrix& out)
  918. {
  919. alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col);
  920. alphaBlend<1, 4>(out.template ref<2, scale - 2>(), col);
  922. alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col);
  923. out.template ref<2, scale - 1>() = col;
  924. }
  926. template <class OutputMatrix>
  927. static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
  928. {
  929. alphaBlend<1, 4>(out.template ref<2, 0>(), col);
  930. alphaBlend<1, 4>(out.template ref<0, 2>(), col);
  931. alphaBlend<3, 4>(out.template ref<2, 1>(), col);
  932. alphaBlend<3, 4>(out.template ref<1, 2>(), col);
  933. out.template ref<2, 2>() = col;
  934. }
  936. template <class OutputMatrix>
  937. static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
  938. {
  939. alphaBlend<1, 8>(out.template ref<1, 2>(), col);
  940. alphaBlend<1, 8>(out.template ref<2, 1>(), col);
  941. alphaBlend<7, 8>(out.template ref<2, 2>(), col);
  942. }
  944. template <class OutputMatrix>
  945. static void blendCorner(uint32_t col, OutputMatrix& out)
  946. {
  947. //model a round corner
  948. alphaBlend<45, 100>(out.template ref<2, 2>(), col); //exact: 0.4545939598
  949. //alphaBlend<14, 1000>(out.template ref<2, 1>(), col); //0.01413008627 -> negligable
  950. //alphaBlend<14, 1000>(out.template ref<1, 2>(), col); //0.01413008627
  951. }
  952. };
  955. struct Scaler4x
  956. {
  957. static const int scale = 4;
  959. template <class OutputMatrix>
  960. static void blendLineShallow(uint32_t col, OutputMatrix& out)
  961. {
  962. alphaBlend<1, 4>(out.template ref<scale - 1, 0>(), col);
  963. alphaBlend<1, 4>(out.template ref<scale - 2, 2>(), col);
  965. alphaBlend<3, 4>(out.template ref<scale - 1, 1>(), col);
  966. alphaBlend<3, 4>(out.template ref<scale - 2, 3>(), col);
  968. out.template ref<scale - 1, 2>() = col;
  969. out.template ref<scale - 1, 3>() = col;
  970. }
  972. template <class OutputMatrix>
  973. static void blendLineSteep(uint32_t col, OutputMatrix& out)
  974. {
  975. alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col);
  976. alphaBlend<1, 4>(out.template ref<2, scale - 2>(), col);
  978. alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col);
  979. alphaBlend<3, 4>(out.template ref<3, scale - 2>(), col);
  981. out.template ref<2, scale - 1>() = col;
  982. out.template ref<3, scale - 1>() = col;
  983. }
  985. template <class OutputMatrix>
  986. static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
  987. {
  988. alphaBlend<3, 4>(out.template ref<3, 1>(), col);
  989. alphaBlend<3, 4>(out.template ref<1, 3>(), col);
  990. alphaBlend<1, 4>(out.template ref<3, 0>(), col);
  991. alphaBlend<1, 4>(out.template ref<0, 3>(), col);
  992. alphaBlend<1, 3>(out.template ref<2, 2>(), col); //[!] fixes 1/4 used in xBR
  993. out.template ref<3, 3>() = out.template ref<3, 2>() = out.template ref<2, 3>() = col;
  994. }
  996. template <class OutputMatrix>
  997. static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
  998. {
  999. alphaBlend<1, 2>(out.template ref<scale - 1, scale / 2 >(), col);
  1000. alphaBlend<1, 2>(out.template ref<scale - 2, scale / 2 + 1>(), col);
  1001. out.template ref<scale - 1, scale - 1>() = col;
  1002. }
  1004. template <class OutputMatrix>
  1005. static void blendCorner(uint32_t col, OutputMatrix& out)
  1006. {
  1007. //model a round corner
  1008. alphaBlend<68, 100>(out.template ref<3, 3>(), col); //exact: 0.6848532563
  1009. alphaBlend< 9, 100>(out.template ref<3, 2>(), col); //0.08677704501
  1010. alphaBlend< 9, 100>(out.template ref<2, 3>(), col); //0.08677704501
  1011. }
  1012. };
  1015. struct Scaler5x
  1016. {
  1017. static const int scale = 5;
  1019. template <class OutputMatrix>
  1020. static void blendLineShallow(uint32_t col, OutputMatrix& out)
  1021. {
  1022. alphaBlend<1, 4>(out.template ref<scale - 1, 0>(), col);
  1023. alphaBlend<1, 4>(out.template ref<scale - 2, 2>(), col);
  1024. alphaBlend<1, 4>(out.template ref<scale - 3, 4>(), col);
  1026. alphaBlend<3, 4>(out.template ref<scale - 1, 1>(), col);
  1027. alphaBlend<3, 4>(out.template ref<scale - 2, 3>(), col);
  1029. out.template ref<scale - 1, 2>() = col;
  1030. out.template ref<scale - 1, 3>() = col;
  1031. out.template ref<scale - 1, 4>() = col;
  1032. out.template ref<scale - 2, 4>() = col;
  1033. }
  1035. template <class OutputMatrix>
  1036. static void blendLineSteep(uint32_t col, OutputMatrix& out)
  1037. {
  1038. alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col);
  1039. alphaBlend<1, 4>(out.template ref<2, scale - 2>(), col);
  1040. alphaBlend<1, 4>(out.template ref<4, scale - 3>(), col);
  1042. alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col);
  1043. alphaBlend<3, 4>(out.template ref<3, scale - 2>(), col);
  1045. out.template ref<2, scale - 1>() = col;
  1046. out.template ref<3, scale - 1>() = col;
  1047. out.template ref<4, scale - 1>() = col;
  1048. out.template ref<4, scale - 2>() = col;
  1049. }
  1051. template <class OutputMatrix>
  1052. static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
  1053. {
  1054. alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col);
  1055. alphaBlend<1, 4>(out.template ref<2, scale - 2>(), col);
  1056. alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col);
  1058. alphaBlend<1, 4>(out.template ref<scale - 1, 0>(), col);
  1059. alphaBlend<1, 4>(out.template ref<scale - 2, 2>(), col);
  1060. alphaBlend<3, 4>(out.template ref<scale - 1, 1>(), col);
  1062. out.template ref<2, scale - 1>() = col;
  1063. out.template ref<3, scale - 1>() = col;
  1065. out.template ref<scale - 1, 2>() = col;
  1066. out.template ref<scale - 1, 3>() = col;
  1068. out.template ref<4, scale - 1>() = col;
  1070. alphaBlend<2, 3>(out.template ref<3, 3>(), col);
  1071. }
  1073. template <class OutputMatrix>
  1074. static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
  1075. {
  1076. alphaBlend<1, 8>(out.template ref<scale - 1, scale / 2 >(), col);
  1077. alphaBlend<1, 8>(out.template ref<scale - 2, scale / 2 + 1>(), col);
  1078. alphaBlend<1, 8>(out.template ref<scale - 3, scale / 2 + 2>(), col);
  1080. alphaBlend<7, 8>(out.template ref<4, 3>(), col);
  1081. alphaBlend<7, 8>(out.template ref<3, 4>(), col);
  1083. out.template ref<4, 4>() = col;
  1084. }
  1086. template <class OutputMatrix>
  1087. static void blendCorner(uint32_t col, OutputMatrix& out)
  1088. {
  1089. //model a round corner
  1090. alphaBlend<86, 100>(out.template ref<4, 4>(), col); //exact: 0.8631434088
  1091. alphaBlend<23, 100>(out.template ref<4, 3>(), col); //0.2306749731
  1092. alphaBlend<23, 100>(out.template ref<3, 4>(), col); //0.2306749731
  1093. //alphaBlend<8, 1000>(out.template ref<4, 2>(), col); //0.008384061834 -> negligable
  1094. //alphaBlend<8, 1000>(out.template ref<2, 4>(), col); //0.008384061834
  1095. }
  1096. };
  1097. }
  1100. void xbrz::scale(size_t factor, const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, const xbrz::ScalerCfg& cfg, int yFirst, int yLast)
  1101. {
  1102. switch (factor)
  1103. {
  1104. case 2:
  1105. return scaleImage<Scaler2x>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
  1106. case 3:
  1107. return scaleImage<Scaler3x>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
  1108. case 4:
  1109. return scaleImage<Scaler4x>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
  1110. case 5:
  1111. return scaleImage<Scaler5x>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
  1112. }
  1113. assert(false);
  1114. }
  1117. bool xbrz::equalColorTest(uint32_t col1, uint32_t col2, double luminanceWeight, double equalColorTolerance)
  1118. {
  1119. return equalColor(col1, col2, luminanceWeight, equalColorTolerance);
  1120. }
  1123. void xbrz::nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight, int srcPitch,
  1124. uint32_t* trg, int trgWidth, int trgHeight, int trgPitch,
  1125. SliceType st, int yFirst, int yLast)
  1126. {
  1127. if (srcPitch < srcWidth * static_cast<int>(sizeof(uint32_t)) ||
  1128. trgPitch < trgWidth * static_cast<int>(sizeof(uint32_t)))
  1129. {
  1130. assert(false);
  1131. return;
  1132. }
  1134. switch (st)
  1135. {
  1136. case NN_SCALE_SLICE_SOURCE:
  1137. //nearest-neighbor (going over source image - fast for upscaling, since source is read only once
  1138. yFirst = std::max(yFirst, 0);
  1139. yLast = std::min(yLast, srcHeight);
  1140. if (yFirst >= yLast || trgWidth <= 0 || trgHeight <= 0) return;
  1142. for (int y = yFirst; y < yLast; ++y)
  1143. {
  1144. //mathematically: ySrc = floor(srcHeight * yTrg / trgHeight)
  1145. // => search for integers in: [ySrc, ySrc + 1) * trgHeight / srcHeight
  1147. //keep within for loop to support MT input slices!
  1148. const int yTrg_first = ( y * trgHeight + srcHeight - 1) / srcHeight; //=ceil(y * trgHeight / srcHeight)
  1149. const int yTrg_last = ((y + 1) * trgHeight + srcHeight - 1) / srcHeight; //=ceil(((y + 1) * trgHeight) / srcHeight)
  1150. const int blockHeight = yTrg_last - yTrg_first;
  1152. if (blockHeight > 0)
  1153. {
  1154. const uint32_t* srcLine = byteAdvance(src, y * srcPitch);
  1155. uint32_t* trgLine = byteAdvance(trg, yTrg_first * trgPitch);
  1156. int xTrg_first = 0;
  1158. for (int x = 0; x < srcWidth; ++x)
  1159. {
  1160. int xTrg_last = ((x + 1) * trgWidth + srcWidth - 1) / srcWidth;
  1161. const int blockWidth = xTrg_last - xTrg_first;
  1162. if (blockWidth > 0)
  1163. {
  1164. xTrg_first = xTrg_last;
  1165. fillBlock(trgLine, trgPitch, srcLine[x], blockWidth, blockHeight);
  1166. trgLine += blockWidth;
  1167. }
  1168. }
  1169. }
  1170. }
  1171. break;
  1173. case NN_SCALE_SLICE_TARGET:
  1174. //nearest-neighbor (going over target image - slow for upscaling, since source is read multiple times missing out on cache! Fast for similar image sizes!)
  1175. yFirst = std::max(yFirst, 0);
  1176. yLast = std::min(yLast, trgHeight);
  1177. if (yFirst >= yLast || srcHeight <= 0 || srcWidth <= 0) return;
  1179. for (int y = yFirst; y < yLast; ++y)
  1180. {
  1181. uint32_t* trgLine = byteAdvance(trg, y * trgPitch);
  1182. const int ySrc = srcHeight * y / trgHeight;
  1183. const uint32_t* srcLine = byteAdvance(src, ySrc * srcPitch);
  1184. for (int x = 0; x < trgWidth; ++x)
  1185. {
  1186. const int xSrc = srcWidth * x / trgWidth;
  1187. trgLine[x] = srcLine[xSrc];
  1188. }
  1189. }
  1190. break;
  1191. }
  1192. }
comments powered by Disqus